Blood-Based Digital Biomarkers and Artificial Intelligence for Early Detection of Alzheimer’s Disease: A Systematic Review

Early detection of Alzheimer’s disease (AD) is crucial for enabling effective preventive and therapeutic strategies. As one of the most prevalent chronic conditions in older adults, AD affects a substantial proportion of the aging population, making timely diagnosis essential for managing its growing global burden. It is also a leading cause of dementia worldwide, underscoring the urgent need for accurate and early diagnostic methods to improve clinical outcomes and disease management¹.

AD was first identified in 1906 by German psychiatrist and neuropathologist Alois Alzheimer, who described unusual memory impairment along with characteristic brain changes in a patient¹.

Alzheimer’s disease (AD) is a progressive, irreversible neurodegenerative disorder and the most common cause of dementia worldwide, characterized by gradual decline in memory, cognition, language, and executive functions. It mainly affects older adults and has a long preclinical phase that may begin decades before symptoms.²

Pathologically, AD is marked by extracellular amyloid-β (Aβ) plaques and intracellular hyperphosphorylated tau neurofibrillary tangles, leading to synaptic dysfunction, neuronal loss,

brain atrophy, and cognitive decline. It is also a multifactorial disorder involving neuroinflammation, oxidative stress, mitochondrial dysfunction, and vascular impairment.

1.1 The Global Burden of AD:

The global prevalence of Alzheimer’s disease (AD) is projected to increase nearly fourfold by 2050, with approximately one in 85 individuals affected worldwide³. In the United States, the proportion of adults with Alzheimer’s disease and related dementias is expected to rise from 1.6% in 2014 to 3.3% by 2060⁴. Advancing age remains the most significant risk factor, followed by the presence of the apolipoprotein E-4 allele, female sex, a history of head trauma, and various cerebrovascular risk factors⁵. The characteristic neuropathological features of AD include amyloid plaques, neurofibrillary tangles, and progressive neuronal loss, predominantly impacting the limbic system and cholinergic neurotransmission pathways⁵.

The growing prevalence of AD presents substantial challenges for healthcare systems and caregivers worldwide^6,7. However, even modest improvements in therapeutic or preventive interventions that delay disease onset and progression by just one year could substantially reduce the global burden of AD, with estimates suggesting the potential prevention of up to 9.2 million cases by 2050³.

As the global population continues to age, Alzheimer’s disease (AD) has become increasingly common and severe, creating a major global public health challenge⁸. The World Health Organization estimates that more than 55 million people are currently living with dementia, and AD accounts for around 60%–70% of all cases⁹. It is mainly characterized by gradual decline in memory, thinking ability, and mood. Because its causes are complex and not yet fully understood, no curative treatment is available, which places a heavy emotional and financial burden on patients, families, and healthcare systems¹⁰. This makes early detection extremely important, as it can help slow disease progression, extend the time for intervention, and improve quality of life.

At present, AD diagnosis mainly relies on neuropsychological tests and biomarker-based methods under the NIA-AA ATN framework¹¹. However, these approaches are often costly, invasive, and difficult to apply on a large scale. For example, “A” biomarkers require cerebrospinal fluid analysis through lumbar puncture or amyloid PET scans¹². Neuropsychological tests are easier to perform but can be subjective and influenced by both patients and examiners, and single-time assessments may sometimes lead to incorrect or missed diagnoses.

To overcome these limitations, researchers are exploring other options such as eye-based biomarkers?, fluid-based markers, and blood-based biomarkers¹³. Even so, these methods still face challenges in repeated testing and long-term monitoring. In this context, digital biomarkers have gained increasing attention. The U.S. Food and Drug Administration (FDA) defines digital biomarkers as “a characteristic or set of characteristics collected through digital health technologies, which serve as indicators of normal biological processes, pathogenic processes, or responses to exposure or interventions, including therapeutic interventions”¹⁴.

In AD, digital biomarkers include objective data collected from devices such as sensors, wearables, and implantable technologies. These include measures like walking patterns¹⁵, eye movements, and speech features. They are non-invasive, objective, and can capture real-life changes in patients’ daily activities, allowing continuous monitoring over time. This makes them highly valuable for early detection and better disease management.

Although several reviews have already examined digital biomarkers in AD, most focus on specific areas such as wearable-based dementia phenotypes or home-based monitoring in mild cognitive impairment (MCI) and early AD¹⁶. They do not fully cover the broader development of the field and need updating due to recent progress. With the rapid rise of artificial intelligence (AI) in healthcare, it is also important to explore how AI can enhance digital biomarkers, although most current reviews focus mainly on traditional biomarkers¹⁷.

This study aims to fill these gaps by providing a systematic review of digital biomarkers in AD, with a focus on AI integration and interdisciplinary research. It aims to (1) identify new research trends, (2) evaluate strengths and limitations of current methods, (3) highlight successful collaboration models, and (4) offer practical recommendations for clinical use.

Fig No. 1: Representation of Risk Factors of Alzheimer^’s Disease

1. 2. Evolution of AI and Neural Networks in Modern Medicine:

Neural networks and artificial intelligence (AI) are increasingly being recognized as transformative technologies for the early detection of diseases and the advancement of modern medicine, significantly reshaping multiple dimensions of healthcare¹⁸. These technologies demonstrate substantial potential in disease prediction, risk stratification, and diagnostic assistance^18,19. Artificial Neural Networks (ANNs) have been extensively applied in areas including cardiology, diagnostics, medical imaging, and radiology¹⁹. In particular, convolutional neural networks, auto-encoders, and recurrent neural networks are employed for applications such as medical image segmentation and arrhythmia detection²⁰. Neural networks possess a strong capability to identify intricate relationships within large and complex datasets, rendering them highly effective for analyzing the extensive diagnostic information produced by contemporary medical technologies²¹. Although clinical expertise continues to play an essential role, neural networks are progressively supporting diagnostic and clinical decision-making processes, with the potential to enhance both accuracy and efficiency in healthcare systems^20,21. Fourcade and Khonsari performed a systematic review²² examining the application of convolutional neural networks (CNNs) in medical image analysis. Their findings indicated that CNNs can enhance and complement the performance of healthcare professionals, especially in visually oriented disciplines such as radiology and pathology. Nevertheless, concerns remain regarding the interpretability of AI-based systems and the necessity for additional investigations to establish their reliability¹⁸.

Fig No. 2: Representation of AI and Neural Network in Modern Medicine.

1.3 Neuropathology of AD:

1.3.1 Deposition of Aβ in Amyloid Plaques:

Deposition of Aβ in amyloid plaques constitutes a central neuropathological feature of Alzheimer’s disease (AD), which is characterized by progressive brain shrinkage accompanied by cortical thinning and cerebral atrophy. The principal microscopic hallmarks of AD include neuritic amyloid plaques and neurofibrillary tangles (NFTs). Amyloid plaques arise from the extracellular accumulation of Aβ peptides generated through the proteolytic processing of amyloid precursor protein (APP) by α- or β-secretase followed by γ-secretase cleavage, resulting in Aβ peptides of different lengths. Although APP metabolism produces greater amounts of Aβ40 compared with Aβ42, the latter exhibits higher amyloidogenic potential because of its hydrophobic C-terminal amino acid residues, rendering it more susceptible to aggregation and plaque formation. In addition to parenchymal deposition, Aβ accumulates within cerebral blood vessels, contributing to cerebral amyloid angiopathy (CAA), in which Aβ40 is the predominant species deposited in vascular walls. Amyloid plaques evolve from diffuse extracellular deposits into mature dense-core plaques containing dystrophic neurites and are detectable using Thioflavin-S staining as well as Aβ immunohistochemical techniques.

Amyloid accumulation follows a characteristic region-specific progression pattern²³, initially appearing within neocortical regions (Stage A), subsequently extending to the hippocampus and broader cortical territories (Stage B), and eventually involving extensive neocortical and subcortical structures such as the striatum, thalamus, and cerebellum (Stage C). Advanced phases of deposition further affect limbic and brainstem regions, including the basal forebrain, substantia nigra, locus coeruleus, and cerebellar structures²⁴. Nevertheless, the extent of amyloid deposition demonstrates only a weak correlation with the severity and duration of dementia^25,26,27, suggesting that although Aβ accumulation represents a defining pathological hallmark of AD, it alone cannot fully account for disease progression.

1.3.2 Deposition of Tau and Phosphorylated Tau:

Pre-tangles and NFTs:

In contrast to amyloid plaques, neurofibrillary tangles (NFTs) exhibit a stronger association with the severity and duration of Alzheimer’s disease (AD)^25,26,27. NFTs are predominantly composed of abnormally misfolded and hyperphosphorylated tau protein, a microtubule-associated protein²⁸. Tau pathology initially develops intracellularly in the form of pre-tangles and is additionally observed within neuropil threads distributed throughout neuronal processes.

The development of NFTs occurs through three distinct morphological stages: (1) pre-tangles, characterized by diffuse and non-fibrillar tau aggregates; (2) mature intracellular filamentous NFTs consisting of aggregated tau and phosphorylated tau; and (3) extracellular “ghost” NFTs, which represent residual structures of degenerated neurons²³. The anatomical distribution of NFTs follows a characteristic spatiotemporal progression described by Braak staging, initially affecting the transentorhinal and entorhinal cortex, subsequently extending to the hippocampus and limbic regions, and ultimately involving widespread neocortical areas while relatively sparing the primary sensory and motor cortices²³. In advanced stages, tau pathology further extends into subcortical regions including the basal forebrain, thalamus, and brainstem nuclei.

Early accumulation of phosphorylated tau (p-tau) has been identified in neurons of the locus coeruleus, with subsequent involvement of the raphe nuclei and basal forebrain occurring prior to cortical pathology, suggesting that tau-related neurodegeneration may originate within subcortical projection neurons several decades before the clinical onset of dementia^23,29. Nevertheless, the precise site of initiation remains controversial, as substantial evidence also supports early pathological involvement of the transentorhinal and entorhinal cortex. Although these observations support the concept of a pathological continuum from pre-tangles to mature NFTs during AD progression, the predominantly cross-sectional design of autopsy-based investigations restricts definitive interpretation regarding temporal causality.  

Fig No. 3: Representation of Neuropathology of Alzheimer^’s Disease.

2. Key Digital Blood Biomarkers for AD Diagnosis:

2.1 Amyloid Biomarkers:

Amyloid Precursor Protein (APP):

Amyloid precursor protein (APP) undergoes proteolytic metabolism through the action of three enzymes known as α-, β-, and γ-secretases³⁰. The generation of amyloid beta (Aβ) peptides occurs through sequential cleavage by β- and γ-secretases, resulting in the release of a soluble extracellular fragment termed sAPPβ. Alternatively, APP can undergo non-amyloidogenic processing via α-secretase cleavage, which produces a longer soluble extracellular fragment known as sAPPα that plays an important role in neuronal growth and synaptogenesis³¹. Among the limited studies investigating blood sAPP concentrations, one study reported no significant differences in plasma sAPPα levels among individuals with mild cognitive impairment (MCI), Alzheimer’s disease (AD), and healthy controls, while plasma sAPPβ levels demonstrated a positive correlation with Mini-Mental State Examination (MMSE) scores³².

A β peptides:

The involvement of Aβ peptides in the pathogenesis of Alzheimer’s disease (AD) is explained by the amyloid cascade hypothesis, which proposes that the disease process is initiated by the excessive production, accumulation, and subsequent aggregation of Aβ into amyloid plaques. Progressive cerebral accumulation of Aβ begins several decades prior to the appearance of cognitive symptoms and constitutes one of the principal pathological hallmarks of the disease. Quantification of the predominant Aβ1–42 and Aβ1–40 peptides in cerebrospinal fluid (CSF) has become an essential component of the updated diagnostic criteria for AD because of their strong correlation with neuropathological observations and positron emission tomography (PET)-based assessments of amyloid pathology³³. Furthermore, measurement of Aβ peptides may provide valuable information for the identification of preclinical AD, even before amyloid deposition becomes detectable through PET imaging. Both Aβ1–42 and Aβ1–40 peptides are considered critical biomarkers for AD diagnosis, and particularly, the Aβ1–42/Aβ1–40 ratio has demonstrated considerable clinical utility as a practical diagnostic indicator that minimizes discrepancies in biologically ambiguous cases and enhances diagnostic confidence for AD^34,35.

A β oligomers (OA β):

In addition to monomeric Aβ peptides, several studies have investigated the detection of plasmatic Aβ oligomers (OAb), which have recently been identified as potential markers of abnormal amyloid status. These oligomeric species have been shown to negatively correlate with Mini-Mental State Examination (MMSE) scores as well as cerebrospinal fluid (CSF) Aβ1–42 levels in patients with Alzheimer’s disease (AD). Various assays demonstrating good clinical performance in certain studies, along with the potential for high-throughput application, have been reported. However, similar to the detection of OAb in CSF using assays that have not yet been incorporated into routine clinical practice, it remains uncertain whether these approaches can surpass the diagnostic performance of monomeric Aβ peptide measurements or achieve sufficient robustness for widespread clinical implementation³⁶.

2.2 Tau Biomarkers:

Total tau (t-tau):

Tau protein is encoded by the MAPT gene and is predominantly expressed in the central nervous system (CNS), primarily within neurons and glial cells. Its physiological roles are diverse and include the stabilization and assembly of microtubules, regulation of axonal transport, and support of axonal growth. Under normal physiological conditions, tau phosphorylation serves as a regulatory mechanism controlling its binding affinity to microtubules. However, excessive phosphorylation of tau promotes its aggregation, leading to the formation of neurofibrillary tangles, a process observed both in normal aging and in pathological conditions such as tauopathies³⁷.

Cerebrospinal fluid (CSF) total tau (t-tau) levels are increased in prodromal Alzheimer’s disease (AD) and AD dementia, although similar elevations are also observed in other neurological conditions including traumatic brain injury and Creutzfeldt–Jakob disease (CJD)^38,39, thereby limiting its specificity as an AD biomarker. Furthermore, plasma and CSF t-tau concentrations show poor correlation with each other. While plasma t-tau is associated with cognitive decline, there is substantial overlap in its levels among different patient groups, which restricts its diagnostic utility in the clinical assessment of AD⁴⁰.

Phosphorylated tau (p-tau):

Cerebrospinal fluid (CSF) phosphorylated tau at position 181 (p-tau (181)) is markedly elevated in prodromal Alzheimer’s disease (AD) and AD dementia, and may also show increases in certain other tauopathies such as progressive supranuclear palsy and frontotemporal lobar degeneration, which reduces its overall specificity. However, levels observed within the frontotemporal lobar degeneration spectrum generally remain below established AD diagnostic cut-offs⁴¹. CSF and plasma p-tau (181) demonstrate a strong correlation, and plasma p-tau (181) is consistently associated with both CSF Aβ biomarkers and tau PET imaging, suggesting that this isoform is more closely linked to AD pathophysiology than total tau (t-tau)⁴². Furthermore, plasma p-tau (181) has been shown to effectively distinguish AD patients from individuals with other tauopathies and to predict the presence of either normal or pathological CSF biomarker profiles⁴³.

In addition, other phosphorylated tau epitopes associated with AD have recently been identified, including p-tau (217) and p-tau (231). These isoforms have been reported to correlate with the presence of amyloid plaques at both asymptomatic and symptomatic stages, although their diagnostic performance varies across studies compared with p-tau (181)⁴⁴. Notably, the increase in plasma p-tau (217) observed in amyloid-positive individuals is greater than that of p-tau (181)⁴², suggesting superior diagnostic accuracy for AD. Moreover, CSF p-tau (217) has been reported to demonstrate higher diagnostic performance for AD compared with CSF p-tau (181) and p-tau (231), although equivalent comparative evaluations in plasma remain to be fully established. Conversely, plasma p-tau (231) has been shown to facilitate identification of both clinical stages and neuropathological features of AD with a diagnostic performance comparable to p-tau (181)⁴⁵.

2.3 Neuronal Damage Biomarkers:

Neurofilament Light Chain Protein (NfL):

Neurofilament light chain protein (NfL) is a cytoskeletal protein that is specifically expressed in neurons and is released into the cerebrospinal fluid (CSF) and bloodstream following neuronal injury. CSF concentrations of NfL have been reported to increase across a broad spectrum of neurodegenerative disorders and are associated with cognitive impairment in patients with Alzheimer’s disease (AD), frontotemporal lobar degeneration, as well as with AD-related pathology observed in dementia with Lewy bodies (DLB) and Down syndrome (DS) patients⁴⁶. Furthermore, plasma and CSF NfL levels demonstrate a strong correlation, and blood-based NfL measurements reflect the overall extent of neuronal damage, thereby enabling monitoring across multiple neurological conditions, although blood NfL is not disease-specific.

A particularly pronounced elevation of NfL is observed in amyotrophic lateral sclerosis, where it serves as an important diagnostic and prognostic biomarker⁴⁷, and its blood-based assay is already routinely implemented in several clinical centers for this purpose. In addition, blood NfL has been reported to reflect white matter abnormalities in patients with mild cognitive impairment (MCI), suggesting its utility as an indicator of microstructural brain alterations. Overall, findings from multiple studies consistently indicate elevated plasma NfL levels in AD patients, supporting its potential diagnostic relevance in Alzheimer’s disease dementia⁴⁸.

S100b and Neuron-Specific Enolase (NSE):

S100B and neuron-specific enolase (NSE) are brain-derived proteins that have been widely investigated as peripheral biochemical markers of brain injury. Several studies have proposed a potential role for S100B as a biomarker in Alzheimer’s disease (AD), as its levels have been shown to correlate with brain atrophy in AD patients, to inversely correlate with Mini-Mental State Examination (MMSE) scores, and to be reduced in the serum of AD patients compared with healthy individuals. In contrast, no significant differences in serum NSE levels have been consistently observed between AD patients and controls; however, a decrease in NSE levels has been reported in AD patients exhibiting more pronounced brain atrophy⁴⁹.

Additionally, astrocytic overexpression of S100B has been linked to elevated levels of interleukin-1 (IL-1) released by activated microglia in AD, suggesting that S100B may reflect underlying inflammatory and pathological processes occurring during disease progression. Nevertheless, interpretation of serum S100B findings should be made cautiously, as this protein is also expressed in multiple peripheral tissues, and observed variations may therefore reflect extracerebral or peripheral processes, such as vascular alterations, rather than AD-specific neuropathology alone⁵⁰.

2.4 Biomarkers of Neuroinflammation:

Glial Fibrillary Acidic Protein (GFAP):

Glial fibrillary acidic protein (GFAP) is an intermediate filament protein predominantly expressed in mature astrocytes located within both gray and white matter regions of the brain, as well as in the cerebellum, subventricular zone, and subgranular zone. In addition to its central nervous system expression, GFAP is also present in peripheral tissues, including Schwann cells, mature enteric glial cells, hepatic stellate cells, and other non-neural cell types. As a marker of astrocytic activation, GFAP is considered reflective of neuroinflammatory processes, and elevated blood levels of this protein have been reported in individuals with preclinical Alzheimer’s disease (AD), indicating its potential utility as an early-stage biomarker⁵¹.

Furthermore, plasma GFAP concentrations have been shown to be associated with both clinical diagnosis and amyloid-β (Aβ) status, with increased levels observed in AD patients across early and late stages of pathology compared with control individuals, supporting its relevance as a biomarker of disease progression. Notably, plasma GFAP levels tend to be higher in patients with more advanced clinical stages compared to those with earlier stages of disease, whereas cerebrospinal fluid (CSF) GFAP levels do not show significant variation across the AD continuum. In addition, a positive correlation has been reported between plasma GFAP levels and tau positron emission tomography (PET) measures⁵².

Triggering Receptor Expressed on Myeloid Cells 2 (TREM2):

Numerous studies have demonstrated that microglial cells, as the primary immune cell population of the central nervous system (CNS), play a crucial role in regulating physiological processes such as programmed cell death, clearance of cellular debris, synaptic pruning, and the development of synaptic connectivity. Alterations in microglial function during aging are considered to be critical determinants in the onset and progression of neurodegenerative disorders, including Alzheimer’s disease (AD). In this context, Aβ peptides have been shown to activate microglial cells, thereby triggering an inflammatory cascade that contributes to neurodegeneration⁵³. Conversely, microglia also possess protective functions, including the phagocytic clearance and degradation of Aβ as well as the production of anti-inflammatory cytokines, which together help to modulate and counterbalance inflammatory responses within the CNS⁵³.

Genome-wide association studies (GWAS) have identified several AD susceptibility genes expressed in microglia, notably the triggering receptor expressed on myeloid cells 2 (TREM2), which is strongly associated with an increased risk of AD. Dysregulation of TREM2 signaling in AD has been linked to impaired microglial phagocytic activity and altered cytokine production, resulting in a functional shift toward a neurodegenerative microglial phenotype⁵⁴.

YKL-40:

YKL-40 (also known as chitinase 3-like 1, CHI3L1) is an inflammatory biomarker that has recently been proposed as a potential indicator for early mild cognitive impairment (MCI) diagnosis⁵⁵. This acute-phase glycoprotein is secreted into the extracellular milieu in response to pro-inflammatory cytokines such as TNF-α, IFN-γ, IL-1β, and IL-6, and it functions as a regulator of cellular processes including proliferation, adhesion, migration, and differentiation. YKL-40 has been reported to be elevated in the brain, particularly in regions surrounding Aβ plaques, as well as in the cerebrospinal fluid (CSF) of Alzheimer’s disease (AD) patients. Increased levels have also been detected in serum across various clinical conditions, including vascular dementia and mixed dementia⁵⁶.

Furthermore, a significant elevation of serum YKL-40, which correlates with CSF levels of the same marker, has been observed in patients with MCI. Importantly, YKL-40 concentrations appear to increase progressively with disease severity and conversion to AD, supporting its potential role in tracking disease progression. Serum YKL-40 has also been shown to correlate with cognitive decline, as assessed by the Mini-Mental State Examination (MMSE), as well as with total tau (t-tau) concentrations⁵⁵.

Cytokines-Chemokines:

A central event in the pathogenesis of Alzheimer’s disease (AD) is neuroinflammatory activation, characterized by microglial stimulation and enhanced production of pro-inflammatory cytokines⁵³. This neuroinflammatory response exhibits a dual and complex role, as it may be beneficial by facilitating amyloid-β (Aβ) clearance, yet also detrimental through sustained glial activation and excessive release of pro-inflammatory mediators that exert neurotoxic effects on neurons⁵³. Variations in serum concentrations of cytokines, chemokines, and growth factors have been reported in individuals with mild cognitive impairment (MCI) and AD, and several of these molecules have been proposed as components of a peripheral blood biomarker signature, some of which have been further validated by the Alzheimer’s Disease Neuroimaging Initiative (ADNI)⁵⁷.

In particular, tumor necrosis factor-alpha (TNF-α) has been extensively implicated in AD-related neuropathology and is consistently found to be elevated in both plasma and cerebrospinal fluid (CSF) of AD patients. Recent findings have further supported the relevance of this cytokine as a blood-based biomarker of AD, demonstrating a strong correlation between TNF receptor 1 (TNF-R1) and neurofilament light chain (NfL) levels across multiple time points, including baseline, 12 months, and 24 months of follow-up in AD patients⁵⁸.               

Fig No. 4⁵⁹: Overview of the blood biomarkers in Alzheimer’s disease: This schematic illustrates the cerebral origin of key blood-based biomarkers in Alzheimer’s disease (AD), including amyloid-related markers, tau proteins, neuroinflammatory mediators, indicators of neuronal injury, cytokines/chemokines, and other reactive biomolecules, as well as their subsequent passage into the peripheral blood circulation.

3. AI-Assisted Early Detection Models for AD:

This paper provides a comprehensive review of state-of-the-art artificial intelligence (AI) models utilized for Alzheimer’s disease (AD) detection, with particular emphasis on classification methodologies, model performance, and the challenges associated with clinical translation. Accordingly, several recent advancements in AI-based models for AD classification are summarized as follows.

3.1 Convolutional Neural Networks (CNNs)

Convolutional Neural Networks (CNNs) are extensively applied in image-based Alzheimer’s disease (AD) detection owing to their strong capability to efficiently extract hierarchical feature representations. These architectures comprise convolutional layers that learn spatial patterns from imaging data, along with pooling layers that reduce computational complexity while retaining critical information. CNNs have demonstrated significant effectiveness in a range of biomedical imaging tasks, including brain tumor classification, object detection, and image segmentation⁶⁰.

3.2 Long Short-Term Memory (LSTM)

Long Short-Term Memory (LSTM) networks are a specialized form of Recurrent Neural Networks (RNNs) developed for sequential data processing, designed to overcome the limitations of standard RNNs through the incorporation of memory cell structures and gating mechanisms. Within this architecture, the forget gate selectively removes irrelevant information, while the input and output gates regulate the storage and retrieval of relevant data over time. Due to their capability to model long-term temporal dependencies, LSTM networks are highly effective in applications such as speech recognition, machine translation, and stock price prediction.

3.3 Hybrid Models

Hybrid models integrate multiple deep learning paradigms, such as Convolutional Neural Networks (CNNs) combined with attention mechanisms or Recurrent Neural Networks (RNNs), to improve both interpretability and predictive performance⁶¹. These architectures are particularly valuable in complex applications such as medical imaging and disease diagnosis, where achieving a balance between high accuracy and model explainability is essential.

Representative implementations include CNN–RNN hybrid frameworks used for the analysis of time-series MRI data to predict disease progression. In addition, multi-branch hybrid architectures have been developed that integrate CNNs trained on different imaging modalities (e.g., MRI and PET), coupled with a shared attention mechanism, enabling more comprehensive and integrative diagnostic insights⁶².

3.4 Transfer Learning

Transfer learning (TL) refers to the use of pre-trained deep learning models that are initially trained on large-scale datasets and subsequently fine-tuned for specific downstream tasks. Models such as BERT for natural language processing and VGG16 for image-based applications exploit transfer learning to adapt generalized feature representations to domain-specific problems. This approach substantially reduces training time and computational cost while maintaining high performance across tasks such as medical imaging analysis, sentiment classification, and natural language processing.

In the context of Alzheimer’s disease (AD) detection, pre-trained architectures including Inception and ResNet have been effectively adapted for diagnostic purposes. Fine-tuning these networks on medical imaging datasets reduces the dependency on large volumes of labeled data and accelerates model convergence. For instance, integrating CNN-based frameworks with pre-trained VGG16 has demonstrated significant improvements in classification accuracy in Alzheimer’s disease-related studies⁶³.

3.5 Generative Adversarial Network (GAN)

Generative Adversarial Networks (GANs) consist of two competing neural networks, namely a generator and a discriminator, which are trained in an adversarial manner. The generator is responsible for producing synthetic (fake) data, while the discriminator evaluates whether the input data is real or generated. Through this competitive learning process, the generator progressively improves its ability to create highly realistic data that closely resembles the original distribution.

GANs have been widely applied in various domains, including image synthesis, deepfake generation, and data augmentation to enhance the training of other machine learning models⁶⁴.

3.6 Auto-Encoder

An autoencoder is a neural network architecture designed to learn efficient data representations through a process of compression and reconstruction. It consists of two primary components: an encoder, which compresses the input data into a lower-dimensional latent representation, and a decoder, which reconstructs the original input from this compressed representation. Autoencoders are widely applied in tasks such as noise reduction, anomaly detection, and synthetic data generation. Common variants include Variational Autoencoders (VAEs) and Convolutional Autoencoders, which are employed across different domains depending on the nature of the input data⁶⁴.

3.7 Vision Transformer (ViT)

Vision Transformer (ViT) introduces a transformer-based architecture for image analysis, replacing the conventional convolutional paradigm. In this approach, images are divided into fixed-size patches, which are then treated as sequential input tokens to a transformer model. By leveraging self-attention mechanisms, ViTs capture contextual relationships between image patches, enabling effective modeling of global dependencies across the entire image. This design makes them highly suitable for tasks such as object detection, image classification, and segmentation. Their strong scalability allows them to perform particularly well when trained on large datasets; however, they typically require substantial computational resources for optimal performance⁶⁵.

3.8 Deep Neural Network (DNN)

Deep Neural Networks (DNNs) constitute a general framework of neural network architectures composed of multiple processing layers. They typically consist of an input layer, several hidden layers, and an output layer, where neurons in each layer are fully connected to neurons in the subsequent layer. DNNs employ optimization techniques such as gradient descent and backpropagation to iteratively learn patterns and representations from data. Due to their flexible and scalable architecture, DNNs are widely applied across diverse domains, including image recognition and financial modeling, providing a robust foundational structure for a broad range of deep learning applications⁶⁶.

3.9 Attention-Based Models

Attention-based mechanisms, such as dual-attention frameworks, are designed to focus on clinically relevant regions within input data, thereby enhancing detection accuracy and reducing false-positive rates. A commonly used modern attention module is the squeeze-and-excitation (SE) block. Representative approaches include self-attention mechanisms in transformer architectures, as well as spatial and channel attention modules in convolutional settings. These attention models are capable of emphasizing critical regions in brain MRI scans that are most relevant for disease diagnosis, thereby supporting clinicians in interpreting and understanding model decision-making processes⁶⁷.

3.9.1 K-Nearest Neighbors (KNN)

K-Nearest Neighbors (KNN) is a simple, non-parametric supervised learning (SL) algorithm used for both regression and classification tasks. The method operates by identifying the k closest data points to a given query instance based on a distance metric, such as Euclidean distance, and then generating predictions accordingly. For classification problems, the output is determined by majority voting among the nearest neighbors, whereas for regression tasks, the prediction is obtained by averaging their values. Although KNN is straightforward to implement and highly interpretable, it can become computationally intensive for large datasets due to the requirement of calculating distances to all data points. It generally performs well on smaller datasets and when features are properly scaled⁶¹.

3.9.2 Support Vector Machine (SVM)

Support Vector Machine (SVM) is a supervised learning (SL) algorithm primarily used for classification tasks, although it can also be applied to regression problems. SVM functions by identifying an optimal hyperplane that best separates data points belonging to different classes in a high-dimensional feature space. It maximizes the margin between the separating hyperplane and the nearest data points from each class, known as support vectors. Additionally, SVM can address non-linear classification problems through the use of kernel functions, such as the radial basis function (RBF). It is particularly effective in high-dimensional settings; however, it may become computationally intensive when applied to large datasets⁶⁸.

3.9.3 Decision Tree (DT)

Decision Tree (DT) is a tree-structured model used for both regression and classification tasks. It partitions the dataset into subsets based on feature values, constructing a hierarchical structure composed of decision nodes and leaf nodes. Each decision node corresponds to a test on a specific feature, whereas leaf nodes represent the final output or prediction. DT models are highly interpretable and can handle both numerical and categorical data effectively. However, they are prone to overfitting, which can be addressed through techniques such as pruning or by integrating multiple trees within ensemble learning methods⁶⁹.

3.9.4 Ensemble Learning

Ensemble learning (EL) is a powerful machine learning (ML) approach that integrates multiple models to enhance predictive performance by combining several weak learners. Its primary objective is to improve overall model accuracy through aggregation strategies that reduce errors and increase generalization. Common ensemble techniques include bagging, which reduces variance (e.g., Random Forest), boosting, which reduces bias (e.g., AdaBoost and Gradient Boosting), and stacking, which employs a meta-learner to integrate predictions from multiple base models for improved final output. These methods frequently outperform individual models and are widely applied in competitive machine learning environments, such as Kaggle, due to their robustness and effectiveness in mitigating overfitting⁷⁰.

3.9.5 Artificial Neural Network (ANN)

Artificial Neural Networks (ANNs) are computational models inspired by biological neural networks (BNNs). They typically consist of an input layer, one or more hidden layers, and an output layer. Each layer is composed of interconnected nodes, where the connections are assigned weights that are iteratively updated during the training process to minimize prediction error. ANNs are capable of modeling complex, non-linear relationships within data and are widely applied in tasks such as regression, classification, and pattern recognition. However, their effective performance generally requires substantial computational resources as well as large-scale datasets⁷¹.

3.9.6 Naive Bayes (NB)

Naive Bayes (NB) is a probabilistic model based on Bayes’ theorem, which assumes conditional independence among features given the class label. Despite this simplifying “naive” assumption, the model often performs effectively in practical applications such as text classification and spam filtering. Variants of Naive Bayes include Gaussian Naive Bayes, which is used for continuous features, and Multinomial Naive Bayes, which is suitable for count-based data such as word frequencies. The algorithm is computationally efficient, requires relatively small amounts of training data, and is robust to irrelevant features. However, its strong independence assumption can limit performance in datasets where feature dependencies are significant⁷².

3.9.7 Logistic Regression

Linear regression is a fundamental machine learning algorithm developed based on multiple observed data points. It identifies the best-fitting linear relationship among these points, which can then be used for prediction or to understand the relationships between different variables. The key advantages of linear regression include its simplicity, high computational efficiency, and strong interpretability. This model is particularly suitable for analyses involving relatively simple datasets and minimal interference, such as standard solution calibration and quantitative assays of pure substances. In such contexts, it can effectively replace traditional manual linear fitting methods, thereby improving both detection accuracy and computational efficiency⁷³.

3.10 Supervised Learning:

Supervised learning methods are capable of learning functions that map inputs to outputs using labeled training data. In this framework, a training dataset consisting of input–output pairs enables the model to identify underlying patterns and relationships within the data. Typically, datasets are partitioned into training and testing subsets to evaluate model performance and generalizability. Common supervised learning tasks include classification and regression. The following section briefly describes various supervised learning approaches integrated with biosensor technologies for Alzheimer’s disease (AD) diagnosis. Each algorithm possesses distinct strengths and limitations, and the selection of an appropriate method depends on the specific characteristics of the dataset as well as the nature of the problem being addressed.

AdaBoost

AdaBoost, also known as adaptive boosting, constructs a strong predictive model by integrating the outputs of multiple weak learners, typically decision trees. This sequential ensemble method converts weak classifiers into a strong classifier through an iterative weighting strategy. In this approach, each training data point is assigned an initial weight, which is subsequently updated during the learning process based on model performance. Specifically, greater importance is assigned to misclassified samples by increasing their weights, while correctly classified samples are assigned reduced weights. These updated weights are then used to train subsequent models, thereby progressively improving overall predictive accuracy⁷³.

Gradient boosting

Gradient Boosting is an ensemble learning algorithm that can be applied to both classification and regression tasks. It is based on the sequential construction of multiple weak learners, typically decision trees, where each subsequent tree is trained to correct the errors of the preceding models. In this way, the model progressively builds a strong predictive framework by focusing on residual errors from earlier iterations. Unlike AdaBoost, which assigns higher weights to misclassified data points, Gradient Boosting identifies errors through the use of gradient-based optimization. In this approach, all decision trees contribute equally in structure, while the overall learning process is controlled through a learning rate parameter, which regulates the contribution of each tree to prevent overfitting and improve generalization performance⁷³.                

Fig No. 5⁷³: Different AI tools for Early detection of Disease

4. Multiple Applications of AI in Modern Medicine:⁷⁴

Artificial Intelligence (AI) has significantly transformed medical imaging and diagnostic analysis. Its application in Alzheimer’s disease (AD) detection spans clinical, research, and technological domains. In particular, AI-driven models are increasingly utilized across diagnostic, therapeutic, and research frameworks. By leveraging advanced architectures and integrating multimodal datasets, these approaches offer substantial potential to enhance early detection, streamline clinical workflows, and support personalized care for AD patients. The key applications and their impacts are summarized as follows:

1. Early Diagnosis and Risk Assessment:

Deep learning (DL) models facilitate early detection of mild cognitive impairment (MCI), a precursor to AD, using structural and functional brain imaging. This enables timely clinical intervention, potentially slowing disease progression, and supports personalized treatment strategies.

2. Clinical Decision Support Systems (CDSS):

Automated DL-based diagnostic systems assist clinicians in identifying AD from MRI, PET, or CT scans with high accuracy. This reduces diagnostic errors, minimizes inter-observer variability, and decreases clinician workload by automating feature extraction and classification.

3. Neuroimaging Biomarker Identification:

DL approaches enable the identification of imaging biomarkers, such as hippocampal atrophy, associated with AD. This enhances understanding of disease pathology and supports the development of targeted therapeutic strategies.

4. Disease Progression Prediction:

Sequential imaging data analyzed using DL models, including hybrid CNN–RNN architectures, allows prediction of MCI-to-AD conversion. This supports monitoring of disease progression and informs long-term clinical management and research planning.

5. Personalized Treatment Monitoring:

DL-based analysis of neuroimaging and clinical data enables evaluation of treatment efficacy. This allows dynamic adjustment of therapeutic strategies and supports real-time monitoring of patient outcomes.

6. Cognitive Function Prediction:

DL regression models can predict cognitive scores (e.g., MMSE) from imaging and non-imaging data, providing a non-invasive proxy for cognitive assessment and supporting routine clinical evaluation.

7. Drug Development and Clinical Trials:

DL models assist in patient stratification by accurately classifying individuals into healthy, MCI, and AD groups. This reduces cohort heterogeneity, improving the reliability of clinical trials and accelerating drug discovery processes.

8. Public Health Screening:

Large-scale automated screening systems powered by DL and integrated with imaging platforms support early AD detection at population levels. This expands diagnostic access in underserved regions and contributes to epidemiological research.

9. Integration With Wearable Devices:

DL models combined with wearable sensor data (e.g., sleep patterns, physical activity, speech) enable continuous monitoring of cognitive decline. This supports non-invasive, real-time risk assessment and early intervention strategies.

10. Real-Time Diagnosis in Telemedicine:

DL-enabled telemedicine platforms analyze remotely acquired imaging and clinical data to provide diagnostic insights. This improves access to specialized care, particularly in rural or resource-limited settings, and reduces patient travel burden.

11. Cross-Modality Analysis:

DL integrates multimodal imaging data such as MRI and PET to enhance diagnostic accuracy. This approach improves classification performance and provides a more comprehensive understanding of structural and functional brain changes.

12. Educational Tools for Medical Training:

DL-based interpretability tools, such as Grad-CAM and saliency maps, are used for medical education. These tools improve understanding of AD-related imaging features and bridge the gap between theoretical knowledge and clinical practice.

13. Synthetic Data Generation:

Generative Adversarial Networks (GANs) are used to generate synthetic neuroimaging data, addressing limitations of data scarcity and class imbalance. This enhances the robustness and generalizability of DL models.